1.2.0
- New model for mobile mapping LiDAR: LitePT-L Mobile Mapping
(MLS, 5 cm), nine classes, NVIDIA CUDA. Weights download
automatically on first use and are SHA-256 verified.
- Fix a Processing shutdown crash after a first-use model
download by retiring download progress callbacks safely.
- Editable output codes for the mobile mapping classes, saved per
model, with a reset to the defaults (pole-like 64, vehicle 65,
fence/barrier 66 are user-defined codes). Processing parameter
OUTPUT_CODES_JSON.
- The visible name is now LiDAR AI Classifier. The package, saved
settings, model cache and Processing algorithm ID do not change.
The existing models are labelled Airborne and keep their
Processing MODEL indexes (0 and 1); the new model is 2.
- Classified layers are styled with the codes actually written, in
2D and 3D, including merged classes. Fixes the 3D style on QGIS 4.
- Outputs with existing extra attributes get a COPC viewing copy
when they are loaded, for QGIS builds whose PDAL reader rejects
them. The classified LAS/LAZ is not modified.
- LitePT dependency checks are retried after a repair and report
the failing package; "Use GPU" switched off and CPU-only PyTorch
are reported as such.
- RTX 5000 Ada is no longer mistaken for an RTX 50 (Blackwell) card
when the compute capability is unknown.
1.1.2
- Stable release: no longer marked experimental.
- The dependency installer starts its helper programs (portable
Python, uv, nvidia-smi) through Qt's QProcess.
- Errors that are safe to ignore are now logged at debug level,
and internal checks raise explicit errors. The plugin passes
the plugins.qgis.org security scan with no configuration file.
- Cancel during the dependency setup stops the running step at
once on Windows.
1.1.0
- New default model on NVIDIA GPUs: LitePT-L Airborne, a point
transformer with 10 cm voxels and 8 classes. The 3D
SegFormer stays available and is the model on CPU and Apple
Silicon. Model selector in the dock, MODEL parameter in
Processing.
- Model weights download automatically, during setup or on the
first run of a model, and are SHA-256 verified before every use.
- QGIS 4 support, tested on QGIS 4.2 (Qt 6) and 3.44 LTR,
including the first-run setup. (GitHub #6)
- Redesigned dock: model status with a one-click switch when a
model cannot run on the computer, a hint saying what is missing
before Run, and the elapsed time with an Open folder button
when a run ends.
- Safer output: results are written atomically, a batch can no
longer overwrite one of its own inputs, every attribute, VLR and
EVLR is kept, truncated files are refused instead of producing
a shorter output, and a custom label field keeps the raw model
classes (cars, trucks and fences stay distinct).
- Installer: prebuilt packages only, certificate checks always
on, no silent switch from GPU to CPU (a failed GPU setup offers
a one-click CPU install), and Plugins > Aerial LiDAR Classifier
> Repair dependencies. Existing environments are rebuilt once.
- Units: vertical units from GeoTIFF vertical CRS codes;
geographic (degree) files are refused even with a units
override.
- The algorithm is available from qgis_process (command line).
- Optional link to the author's live LiDAR course in the panel,
About and the plugin menu; it can be hidden.
1.0.3
- Fix: the plugin could not open on QGIS builds with Python
3.9 to 3.11 (Ubuntu 22.04, Debian 12, macOS 3.34):
"SyntaxError: unterminated string literal" in venv_manager.py.
All files now compile on Python 3.9+. (GitHub #3)
- Fix: files in feet (most US LiDAR) were fed to the model
unconverted, so buildings came out as Wire-Conductor and
Transmission Tower and the run was about 13x slower. The unit
is now read from the LAS CRS and converted to metres for the
model; a manual override exists in the dock (Advanced
parameters > Input units) and in Processing (UNITS).
(GitHub #5, #2)
- Fix: flat tiles (less than 5.12 m of relief) were classified
almost entirely as Building; the block grid now always covers
the extent and unpredicted points are reported as ASPRS 0.
- Fix: fresh installs could end up with a CPU-only torch because
the second install phase let uv replace the CUDA torch with
PyPI's newest release; the CUDA build is now pinned.
- Fix: certificate verification for the package hosts is on by
default; it is relaxed only after a TLS failure, with a warning.
- Fix: a failed CUDA cascade could leave the venv without torch
while it was reported ready.
- Fix: Apple Silicon crashed at run time ("Torch not compiled
with CUDA enabled"); the device is now passed end to end.
- Fix: the Processing algorithm loaded the result layer from the
worker thread; it is now loaded on completion, and the
classified file is exposed as the OUTPUT_FILE output.
- Fix: version shown in About and the Processing provider.
- Fix: auto tile size on corridor-shaped extents; sliver tiles.
- Streaming mode uses 1 byte per point for predictions instead
of 4.
1.0.2
- Fix: Linux install failed at the venv pre-flight check with
"error while loading shared libraries: libpython3.12.so.1.0:
cannot open shared object file" on every distro that uses
python-build-standalone (i.e. all of them). The
`python -m venv --copies` flag copies the standalone python3
binary into venv/bin/ but does NOT copy the libpython next
to it (venv has no notion that python-build-standalone
ships libpython as a separate file). After the copy, the
binary's RPATH=$ORIGIN/../lib resolves to an empty
venv/lib/ and every invocation dies. `--copies` was added
to dodge Windows AV quarantine of the redirector launcher.
It is now applied only on Windows; Linux and macOS use
symlinks (the default), which point back at the standalone
tree so the RPATH lookup still finds libpython.
- Fix: "Use GPU" checkbox stayed unchecked across QGIS sessions
after a fresh install. The dock's closeEvent persisted the
checkbox state even when the GPU probe had forced it off
(torch not yet importable, or CUDA wheel/driver mismatch
during the cu128 -> cu118 cascade). The False stuck and left
users silently on CPU on the next launch. The plugin now
only persists the GPU preference when the user could
actually choose it (i.e. the checkbox is enabled), and a
one-time settings migration resets the stale False on first
launch so existing v1.0.1 users get GPU back automatically.
- Fix: streaming I/O no longer silently drops the trailing
partial chunk on very large LAS files. v1.0.1 skipped the
last few thousand points whenever laspy hit
"buffer size must be a multiple of element size" at EOF.
The chunk loop now uses read_points(n) directly and falls
back to a raw-byte recovery path that decodes every
whole-point-record left on disk, so the streaming output
has the same point count as the input.
- Polish: trimmed the long About text in the QGIS Plugin
Manager so the rating widget is visible without scrolling.
1.0.1
- Fix: corporate SSL inspection blocking the install. uv now
uses the OS native TLS (--native-tls) so Windows-installed
corporate CAs are trusted, and download.pytorch.org is in
the allow-insecure-host list alongside pypi.org and
files.pythonhosted.org. Same flag added to the pip code path
via --trusted-host.
- Fix: Linux first-install failed at ensurepip step because
python-build-standalone Linux tarballs do not ship the
bundled pip wheel. The venv is now created with --without-pip
(we use uv for all package installs anyway).
- Fix: CUDA wheel selection now cascades through cu128 -> cu126
-> cu124 -> cu121 -> cu118 instead of giving up at the first
driver-version miss. Recovers GPU acceleration for NVIDIA
drivers older than 550.
- Fix: cap torch version per cuda index (cu118/cu121 -> torch
<2.6, cu124 -> torch <2.8) so uv no longer picks the latest
+cpu wheel from the cuda index instead of the latest +cuXXX.
- Add: install marker recording the plugin version that built
the venv. On version mismatch the Setup dock reopens with a
one-click Reinstall that wipes the stale venv automatically.
- Add: classification-coloured 3D renderer auto-attached to
loaded layers so the 3D Map View renders points in 3D out of
the box.
1.0.0
- Initial public release on the QGIS plugin repository
- 3D SegFormer (UrbanFiltering, TreeAIBox / NRCan) integration
- QGIS Processing algorithm + provider (Toolbox, Modeler, qgis_process
CLI)
- Isolated per-user venv dependency installer (PyTorch CPU /
CUDA 12.1 / 12.4 / 12.6 / 12.8), compute-capability- and
driver-version-aware wheel selection
- Background QgsTask with progress and cancellation
- LAS / LAZ / COPC input; LAS or LAZ output (matches input)
- Spatial tiling and streaming I/O for files larger than RAM
- Optional automatic loading of results as point-cloud layers
- SHA-256 verification of downloaded model weights
- Network calls go through QgsBlockingNetworkRequest (proxy /
auth / certificates honoured per QGIS plugin guidelines)

None